ANN Blood Plasma Level

8/3/2019 ANN Blood Plasma Level

1/13

.Decision Support Systems 30 2000 139151

www.elsevier.comrlocaterdsw

Estimating drugrplasma concentration levels by applying neuralnetworks to pharmacokinetic data sets

Kristin M. Tolle a,), Hsinchun Chen a,1, Hsiao-Hui Chow b,2

aMIS Department, Karl Eller Graduate School of Management, Uniersity of Arizona, Tucson, AZ 85721, USA

bCollege of Pharmacy, Uniersity of Arizona, Tucson, AZ 85721, USA

Abstract

Predicting blood concentration levels of pharmaceutical agents in human subjects can be made difficult by missing data

and variability within and between human subjects. Biometricians use a variety of software tools to analyze pharmacokinetic

information in order to conduct research about a pharmaceutical agent. This paper is the comparison between using a

feedforward backpropagation neural network to predict blood serum concentration levels of the drug tobramycin in pediatric

cystic fibrosis and hemotologiconcologic disorder patients with the most commonly used software for analysis of

pharmacokinetics, NONMEMw. Mean squared standard error is used to establish the comparability of the two estimation

methods. The motivation for this research is the desire to provide clinicians and pharmaceutical researchers a cost effective,

user friendly, and timely analysis tool for effectively predicting blood concentration ranges in human subjects. q2000

Published by Elsevier Science B.V.

Keywords: Artificial neural networks; Medical applications; Medical user interfaces; Pharmaceutical estimation applications; Pharmacoki-

netic prediction

1. Introduction

Artificial neural networks, which have strong sta-

tistical prediction capabilities, continue to gain ac-

ceptance as data analysis tools. Testing in this and

other research has shown that neural networks can be

trained to estimate plasma concentration values of

pharmaceutical agents without relying on complex

)

Corresponding author. Tel.: q1-520-621-3927. . E-mail addresses: [email protected] K.M. Tolle ,

[email protected] H. Chen , [email protected] H.-H. Chow .

1Tel.: q1-520-621-4153.

2Tel.: q1-520-626-4055.

computation models andror cumbersome statistical

prediction applications. If the need for complex mod-

eling were eliminated, testing results could be gener-

ated more quickly and easily than it is accomplished

using currently available technology.w xBrier et al. 3 examined the use of neural net-

works for population pharmacokinetic analysis, con-

cluding that NONMEM and the neural networks

provided comparable predictions of plasma drugconcentrations. Our research question was to deter-

mine whether a neural network application specifi-

cally designed for the prediction of blood serum

concentration levels of pharmaceutical drugs could

be an effective replacement for current statistical

analysis methodologies. Our goal was to create an

intelligent tool that could assist clinicians in opti-

0167-9236r00r$ - see front matter q 2000 Published by Elsevier Science B.V. .P I I : S 0 1 6 7 - 9 2 3 6 0 0 0 0 0 9 4 - 4


2/13

( )K.M. Tolle et al.r Decision Support Systems 30 2000 139151140

mally administering drugs and provide pharmaceuti-

cal scientists with valuable data gathered in clinical

settings, more promptly and efficiently.

In this paper, we first present a brief background

description of pharmacokinetic analysis, statistical

prediction methods, and neural networks. Next, we

discuss an experiment in which we compared the

predicative capability of a neural network application

with NONMEM, the industry standard application

currently used for pharmacokinetic data analysis.

The paper then describes usability issues associated

with a proposed neural network application, the user

interface, the types of analysis for which it could be

used to assist medical researchers and clinicians in

their work, and the social implications of adopting

the neural network application in place of NON-

MEM and similar data analysis tools.

1.1. Background

1.1.1. Pharmacokinetic analysis

Pharmacokinetics is the study of how various

biological processes influence the effectiveness of

drugs the quantifying of determinates of drug

concentration.

Information about the proper administration of

pharmaceutical drugs in clinical settings is obtained

from varied sources. Some data are collected follow-

ing a very formal set of procedures. Much of theinformation, however, must be obtained by observ-

ing the actual clinical administering of the drugs.

However, the non-stringent nature by which clinical

treatments are administered results in many of the

data gathered in clinical settings being in a format

which is not easily applicable to standardized statisti-

cal and data analysis.

Although clinical information gathering does not

use rigorous methodology, data collected from clini-

cal studies of pharmaceutic agents are still very

useful to assist clinical physicians in modulatingw xtreatment dosages to patients in their care 13 . Such

data record the administration of a drug to individu-

als and the subsequent observation of drug levels .most often in blood plasma . The study of this type

w xof information is referred to as Pharmacokinetics 2 .

In rigorous methods of gathering population dataon a particular drugs effects, an individual the

.subject contributes one instance of information. In

clinical settings, a patient may contribute multiple

instances of information, potentially occurring over

an extended period of time, thereby creating the need

for an analysis model, which takes into consideration

a time series of responses known as repeated

measures. Related to this is the problem known as

imbalance, in which one patient may contribute a

series of data, while another patient may only pro-

duce one datum. A third issue, confounding, occurs

when a patient is given specific dosages of a particu-

lar drug based on that patients prior reactions, in

contrast to more stringent methods that requirew xdosages to be given randomly 13 . These issues

make it difficult to devise statistical analysis tools to

predict pharmacokinetic parameters such as blood

serum concentration levels, volume of distribution,and clearance how quickly a pharmaceutical agent

.leaves the blood stream .Other data analysis problems stem from popula-

.tion within patient demographics such as age,

weight, seriousness of the illness, gender, etc. Some

demographic data may change during the period

when treatment is administered, further complicating

determination of the proper administration of drug

dosages.

1.1.2. Statistical prediction methodologies

The current standard for the analysis of popula-

tion pharmacokinetic data is the application NON-MEM, developed by the University of California at

San Francisco. NONMEM is a data analysis tech-nique for fitting nonlinear mixed effects statistical

.regression-type models that is mainly applied in the

estimation of pharmacokinetic and pharmacody-w xnamic data 2 . NONMEM users must have high

levels of understanding of statistics and pharmacoki-w xnetics 5 in order to use the application successfully.

Furthermore, a user must initially determine the ap-

propriate statistical model for the NONMEM pro-

gram to use for data analysis before any interactionwith an application is undertaken and this can take

many hours or days to complete.

The UNIX version of NONMEM requires the use .of FORTRAN-like commands Fig. 1 to convey

modeling and data information to the application,

which further complicates NONMEM usage. Only a

small number of researchers in the field of biomet-


3/13

( )K.M. Tolle et al.r Decision Support Systems 30 2000 139151 141

Fig. 1. An example of a NONMEM session.

rics have the skills needed to be proficient at phar-macokinetic data analysis using NONMEM, making

them a scarce and expensive resource.

Several other computerized methods of doing

pharmacokinetic analysis exhibit varying degrees ofw xusability and predictive capabilities, but Roe et al. 9

found NONMEM to be more flexible, have fewer

limitations in modeling of data, and be consistently

more successful at predicting pharmacokinetic pa-

rameters than other similar applications. Since these

are important issues to pharmaceutical researchers,

NONMEM remains the accepted standard for con-

ducting pharmaceutical studies.

1.1.3. Estimation using neural networks

Different models of neural networks exist. Gen-

eral details about artificial neural networks can bew xfound in Ref. 11 . A taxonomy of the types of

w xartificial neural networks can be found in Ref. 7 .

Our design employs a connectionist, feedforwardrbackpropagation neural network.

Since the predictive capability of neural networks

is typically nonlinear, it is appropriate to explain that

feedforward neural networks perform a kind of non-

linear regression in which a multilayer network is

trying to find a low-order representation in the

weights between the network layers. That representa-

tion itself is, in general, a nonlinear function of the

physical input variables that allows for the interac-

tions of relationships among many input variables atw xone time 10 . Thus, the inputs become dependent on

one another through network interaction and ulti-

mately, generate nonlinear estimations as output

variables.

Backpropagation, selected for our design, is a

neural networking algorithm in which activation is

passed forward through the network and the output

unit activations are compared with a teaching vector.


4/13


These represent the inputroutput pairs. The compar-

ison of inputroutput pairs results in error scores,

which are used to propagate changes back down

through the layers of weights. Weights represent the

numerical strength of the connections or links be-

tween a node and its neighbors in a neural networkw x1 and can have either positive or negative valuesw x11 . These weights represent the AintelligenceB of

the network the essence of its predictive capabil-

ity.

The role of an activation function is to combine

the input being broadcast to a node from other nodes

in a network. A typical activation function com-

presses the network activation impinging on a nodew xbetween predetermined limits 1 usually a value

between zero and one. We chose a sigmoidal, or

s-shaped, activation function on the basis of its

excellent predictive capabilities demonstrated in a

previous experiment involving the estimation of toxinw xconcentration in soil samples 4 .

During the learning process, the sigmoidal unit isroughly linear for small weights a net input near

.zero and gets increasingly nonlinear in its response

as it approaches its points of maximum curvature on

either side of the midpoint. Thus, at the beginning of

learning, when weights are small, the system is

mainly linear and seeking a linear solution. As the

weights grow, the network becomes increasingly

nonlinear and begins to move toward a nonlinear

solution to the problem. This linearity property makesthe units more robust and allows the network to

reliably attain the same solution in repeated experi-w xmentation 10 . Thus, two different training sessions,

using the same input data and randomly initialized

weights, should consistently predict the same results.

1.1.4. Neural network parameters

The ability to train multilayer networks is an

important step toward building intelligent applica-

tions. Neural networks must learn their own repre-

sentations because it is not possible to program themw xby hand 8,10 . The optimization of neural network

parameters is critical in order to achieve the best

possible predictive ability.

Five different parameters can be adjusted in the

creation of a backpropagation neural network: hid-

den units, number of layers, learning rate, momen-

tum and number of epochs. The number of hidden

units refers to the number of nodes plus a threshold

node which are to be placed between the input and

output vectors. Layers represent the number of layers

of hidden units between the input and output vectors.

Learning rate is the numeric value by which the

weights between the input, hidden, and output layers

are adjusted. Momentum is a parameter, which can

increase the pace of learning, potentially reducing

the amount of time that it takes to train the network.

The number of epochs refers to the number of times

a data set is applied to the neural network for

training, tuning, and testing.

1.1.5. Currentr possible scenario

For researchers, the greatest problem with the

current system is that they must develop their own

statistical prediction models in order to study a drug. .This is a time-consuming process often days and

takes a high level of skill.Work conducted in research facilities commer-

.cial, private, and academic conduct analysis on

pharmaceutical agents currently is reported to re-

searchers either as information accompanying a drug

or through journal articles and other publications.

Clinicians must rely on the accuracy of such research

although in practise, they develop a AfeelB for the

effectiveness of a drug at a certain blood level

concentrations for specific patients. They therefore

dose drugs based both on past experience and recom-

mended dosages generated by pharmaceutical com-panies and researchers without having any opportu-

nity for discussion of findings. Also, the physicians

data and knowledge may or may not be shared with

colleagues within their facility andror field.

No application exists today which would allow

medical practitioners to quickly and easily adjust

dosage to patients who have differing pharmacoki-

netic parameters, which means they must rely on

documentation and experience to correctly dose a

patient to maintain a plasma concentration level in

an effective range.Providing clinicians with NONMEM would not

necessarily generate more research; the software re-

quires an ability to do extensive statistical modeling

beyond the skills of many clinicians. In contrast, a

neural network works empirically. Once a patients

pharmacokinetic parameters are entered, the best

possible dosing regimen for a particular individual


5/13


can be found simply by using slider bars to adjust

dose and the interval time between dosages. This

application would not only be an effective clinical

tool, but could also facilitate information sharing,

resulting in better care for patients.

Pharmaceutical researchers would also benefit.

Studies could be conducted more quickly than if they

had to develop a prediction model for every study.

Because the neural network would train itself to

maximum prediction efficiency when given a certain

set of parameters, the researcher could generate sev-

eral studies in the amount of time it previously took

to do one.

2. Experimental test bed for population analysis

The data set for our experiment contained infor-

mation regarding tobramycin, a drug used for re-pressing infectious diseases of the blood. Patients

eligible for the study had been diagnosed with either

cystic fibrosis or a hemotologiconcologic disorder,

were less than 18 years of age, had been receiving

tobramycin more than 24 h, and had had peak and

trough concentration blood samples taken at approxi-

mately 2 and 6 h post-initiation of a 30-min drug

infusion of tobramycin.

The data set, collected from 1983 to 1992, consti-

tuted a total of 311 instances of patient information

gathered from 101 patients. Originally, this informa-tion was collected for a study to determine whether

illness had a significant effect on the effectiveness of

the tobramycin. The parameters of the data set are

shown in Fig. 2.

There were two possibilities for data set represen-

tation: if peak and trough concentration levels were

ignored, the data set would contain 622 data points.

If peak and trough concentrations were predicted

separately, the data set would contain 311 data points.

Consistent with most data sets collected in clinical

settings, several patients had contributed a small

number of inputs while others had many inputs over

an extended period of time.

3. The experiment

3.1. NONMEM analysis

The NONMEM analysis used 622 data points and

predicted blood concentration values for each patient

in the NONMEM study without consideration ofpeak and trough values. The pharmacokinetic model

used for estimating blood serum concentration can .be represented in Eq. 1 .

y Cl rV T .i ik 1 y e .i oC s e y Cl rV . 1 . .ti j i i jy Cl rV t .i iCl 1 y e .i

Supporting equations for clearance and volume of . .distribution are found in Eqs. 2 and 3 .

h1Cl s u = ln weight qu =age e 2 . .i 1 i 3 ih1

Vsu =

ln weight e 3 . .i 2 iwhere for child i, C is the observed concentrationi jat measurement j. k is the infusion rate. T is thei o i

Fig. 2. Data set parameters.


6/13


Table 1

Normalization of input parameters

Input parameter Test I Test II

Age agermax. age agermax. age

Sex nra nra

Illness nra nra

Weight weightrmax. wt. weightrmax. wt.

Dose dosermax. dose dosermax. doseInterval between dosages intervalrmax. interval intervalrmax. interval .Time of blood drawn generic Test I only timermax. time .Time of blood drawn peak Test II only timermax. time .Time of blood drawn trough Test II only timermax. time

infusion time. t is the dosing interval. T is thei i jelapsed time at measurement j after the end of

infusion. Cl is the clearance. V is the volume ofi idistribution and ln is the natural log.

3.2. Neural network testing phases

In our initial test, which directly matched the

NONMEM analysis, the original data set was di-

vided into the six demographic and dosage inputs

plus the time of the blood drawing resulting in a total

of 622 data points. A single output was generated

the plasma concentration of tobramycin given at the

time of blood drawn. The network topology was a

seven-inputrone-output neural network with 30 hid-

den units.

In the second test, the original 311 data points

were presented to the neural network. Peak and

trough concentrations were predicted as separate val-

ues. The resulting network topology was eight inputs

and two outputs with 23 hidden units.

In both data sets, two-thirds of the data pointswere used for the training phase of the neural net-

work. The remaining data points were used for the

tuning phase. The input units were normalized as

shown in Table 1 for both tests. This resulted in the

best predictive capability from the neural network.

The output vector of the neural network, a value

between 0 and 1, needed to be modified in order to

make comparisons between the networks output

vector and the actual data. This value was normal-

ized to the same range of values. This was done .using the formula found in Eq. 4 .

ln actual concentration y ln minimum concentration value . .4 .

ln maximumconcentrationvalue y ln minimum concentration value . . .

3.3. Neural network parameters

The goal of selecting the settings for the different

parameters for a neural network is to minimize the

Table 2Testing parameter ranges

Parameter Range Increment

Epochs 010 000 1

Hidden units 360 1

Learning rate 0.050.60 0.05

Momentum 0.050.90 0.05

.mean standard squared error MSSE . While other

parameters are held constant, each parameter is tested

using the data set. The ranges of parameters tested

are shown in Table 2 and the optimal parameters,

Table 3Optimal parameters selected

Test I Test II

Epochs 5400 800

Hidden units 30 23

Learning rate 0.35 0.4

Momentum 0.0 0.0


7/13


Fig. 3. Results of Test I one-way analysis of variance and paired two sample t-test.

which were selected are shown in Table 3. Although

momentum was tested, it was not used in the final

selection of parameters due to increased MSSE in

the results.

3.4. Results of testing

The most direct comparison between the neural

network results and the NONMEM analysis was Test

Fig. 4. Results of Test II one-way analysis of variance and paired two sample t-test.


8/13


I no separation of peak and trough concentra-

tions, 622 data points. The result of these tests

showed that the neural network could predict blood

concentration levels more successfully than NON-

MEM, however, not at a statistically significant level.

The best result was obtained when peak and trough

concentrations were predicted separately in Test II.

In Figs. 3 and 4, created using MiniTabw, the

one-way analysis of variance shows that the neural

network predicted blood serum concentration levels

better than NONMEM. The mean squared error .MSE associated with the prediction was lower in

each of those cases. The p-value for Test I, the most

direct comparison, was approximately 10%. The p- .value for Test II, which represents the lowest best

p-value, was approximately 9%.

4. A medical decision support application for clin-icians and researchers

4.1. Usability issues

In order for a software application to be widely

accepted, regardless of its type, it must be easy to

use. The growing emphasis on usability is one of the

most dominant of the current trends in computingw x6 .

NONMEM has two difficulties: the application

can be complex to use and the development ofprediction models is required. A biometrician de-

scribed the process of model development as taking

a AvariableB length of time, sometimes as long as a

day or several days. Since the cost of a typical

biometricians time is expensive, there would be a

cost benefit if he spent less time modeling and more

time conducting research about gathered data. The

biometrician also found the application difficult to

use both because of the FORTRAN-like commands

and the operating system in which it was used .UNIX .

Timeliness of information: in clinical settings, it is

necessary to have relevant drug information readilyw xavailable 12 . Since a neural network should be able

to generate predictions in a few minutes or seconds,

clinicians could get information instantly, not read

about them later after a biometrician has performed a

study.

Ultimately, the criterion for acceptance is based

on the performance of the application how accu-

rately it works. Proof that the neural network will

consistently predict blood concentration values equal

to or better than currently accepted statistical tech-

niques must be provided to researchers in order to

convince them to change from their proven method-

ologies.

4.2. The data analysis tool

The goal is to create an application that clinicians

and researchers can reliably use to generate concen-tration levels and potentially other parameters as

.well and perform trend analyses. Clinicians would

then have necessary information at their fingertips

and pharmaceutical researchers would have a tool to

make the analysis of data easier.

The proposed interface would address two possi-

ble types of analysis: prediction and data analysis. At

invocation of the application, users could choose

which of these tools they would prefer to use. The

interface will be developed for use on a wide variety

of platforms, using Java as the interface development

tool.

Our proposed data analysis tool was developed in .ANSI C the neural network component , with a

Delphi interface component on Windows NT. Delphi

and NT provide a convenient environment for quick

prototyping. Several medical and pharmacy school

researchers have participated in the design of this .analysis tool Fig. 5 .

4.2.1. Whatif analysis

The Whatif analysis interface would be able to

determine the potential blood concentration level in a

patients blood stream. Two key types of data would

be provided to the neural networks input vector:

patient data and drug application data. Patient data

are parameters such as age, weight, illness, etc. Drug

application data are the clinicians dosing regimen

such as dosage, dosing interval, time of blood drawn,etc.

As previously noted, patient data vary over time.

Patients get older, may gain or lose weight, and

perhaps grow increasingly ill or show signs of recov-

ery. Whatif analysis will be a helpful tool to assist

clinicians in predicting what the plasma concentra-

tion level will be, based on changes.


9/13


Fig. 5. Main screen of pharmacokinetic predictor.

A likely scenario follows: a patient with cystic

fibrosis was admitted to the hospital at the age of

3.25 years. The patient was treated for an infection

with tobramycin. Two years later, the patient reen-ters the hospital for treatment of another infection.

The patients previous data are helpful, but certain

demographics have altered, so reaction to the drug is

likely to be different. The patient is now age five.

Between the ages of 3 and 5, children grow very

rapidly, usually experiencing a considerable weight

gain. The infection is less advanced than it was

during the previous visit. At the minimum, three

parameters have changed. The Whatif analysis can

accept information changes and quickly predict a

blood concentration level for a particular dosing

regimen.

Drug applications can also be dynamic. Perhaps, a

clinician would prefer to see a particular concentra-

tion level in a patient, for instance, the recommended

optimal level. By using the Whatif analysis, the

clinician could alter the dosage or dosing interval to

reach the desired blood serum concentration level in

the patient. This strategy would be particularly useful

if previous dosing information about the patient were .unavailable Fig. 5 .

A likely scenario here might be one in which thepatient had never been previously treated, but is now

entering the hospital with an infectious disorder. The

clinician could enter patient data and within a few

seconds determine, based on that information, what

that patients blood concentration level will be after

administration of a particular dosage andror dosing

interval. Information could be reentered until the

clinician determines the regimen that best fits the

desired concentration level without having to test the

patients reaction to the drug by dosing him or her

and then drawing blood and checking peak and .trough concentrations Fig. 6 .

Graphing ability would be incorporated into the

interface, allowing the clinician to choose a parti-

cular drug application parameter such as dosage,

holding all other parameters constant. The Whatif

analysis would display a chart showing the blood .concentrations for different dosages Fig. 7 .


10/13


Fig. 6. Data parameter selection screen of pharmacokinetic predictor.

Since the data set and the weights associated with

it would already exist, predicting the blood concen-

tration level of the patient would be a quick and easy

process, requiring only a few seconds of CPU pro-

cessing time. At present, no application of this kind

is available to assist clinicians in finding an effective

dose range for an individual patient.

4.2.2. The training session

Targeted toward meeting the requirements of

pharmaceutical researchers, the Training Session in-

terface would be used to present new data sets to the

neural network for analysis or to take an existing

data set and alter the fields included to analyze how

highly each input is correlated with predicting the

output. Additionally, the interface would be useful to

a researcher who might wish to predict peak and

trough concentrations or a single concentration level

either peak, trough or a generic value. The Train-

ing Session would give researchers the flexibility to

set their own input and output vectors, allowing them

to conduct a broad range of research relating to a

particular drug in a short period of time.


11/13


Fig. 7. Graph of weight vs. concentration based on selected parameters.

A possible scenario might involve a pharmaceuti-cal company that would like to determine whether

the type of illness a patient has is highly correlated

with the blood concentration level of a drug. The

data set is loaded into the Training Session and input

and output vectors are established. The neural net-

work is trained the first time, including the illness

parameter and then a second training session is run

with the parameter removed. The Training Session

can present the researcher with error information. If

the error is significantly higher during the second

session, it can be concluded that the illness is likely

to be a highly correlated parameter and that perhaps,

the drug is better for use in treating some illnesses

than others.

An important usability feature of the Training

Session will be the ability to automatically scale and

normalize the data set, select optimal network pa-

rameters, and train the network to predict values,

without the users having to understand the complex-ity of neural networking. No extensive training for

the researcher would be required.

5. Conclusions and future work

The results of the study comparing NONMEM

and neural network show that the neural network has

predictive capability equal to or better than NON-

MEM. Although statistical significance was not es-

tablished, we were able to prove that a neural net-

work can accurately predict blood concentration level

in human subjects making the two methods inter-

changeable tools for effectively estimating concen-

tration levels.

In addition to accuracy, the neural network appli-

cation has the advantage of producing results empiri-

cally, without the need for developing statistical


12/13


prediction models. This puts the power of generating

results in the hands of clinicians who may not be

well trained in this type of analysis methods. Further,

it enables biometricians to have more time to con-

duct analyses. As a cost effective, accurate pharma-

cokinetic estimation application, it should be an ex-

cellent tool for both researchers and practitioners.

The future directions of this project are to com-

plete the proposed modules in the prototype. Once

these are completed, we plan to conduct a usability

study and cost benefit analysis with both clinicians

and researchers to investigate adoption issues and

functionality.

Acknowledgements

This project was supported in part by the follow-

ing grants: Special Information Services, National .Library of Medicine NLM , National Institutes of

.Health NIH , ASemantic Retrieval for Toxicology

and Hazardous Substance DatabasesB, 1996 1997; .National Cancer Institute NCI , National Institutes

.of Health NIH , AInformation Analysis and Visual-

ization for Cancer LiteratureB, 19961997; and AT

&T Foundation Special Purpose Grants in Science

and Engineering, 19941996.

References

w x1 C. Beardon, Artificial Intelligence Terminology: A Reference

Guide, Halstead Press, New York, 1989.w x2 A.J. Boeckmann, L.B. Sheiner, S.L. Beal, NONMEM Users

Guide Part V. Introductory Guide, University of Califor-

nia San Francisco, San Francisco, 1994, pp. 17, Chap. 1.w x3 M.E. Brier, J.M. Zurada, G.R. Aronoff, Neural network

predicted peak and trough gentamicin concentrations, Phar- . .maceutical Research 12 3 1995 .

w x4 H. Chow, H. Chen, T. Ng, Myrdal, S.H. Yalkowski, Using

backpropagation networks for the estimation of aqueous ac-

tivity coefficients of aromatic organic compounds, Journal of

. .Chemical Information Computer Science 35 4 1995 .w x5 L. Kilmartin, E. Ambikairajah, S.M. Lavelle, A neural net-

work diagnostic tool for liver disorders. Paper presented at

Neural Computing Research and Applications: Part Two,

1992.w x6 R. Kling, Organizational analysis in computer science, The

. .Information Society 9 2 1993 7187.w x7 R.P. Lippmann, An introduction to computing with neural

. .nets, IEEE ASSP 4 2 1987 422.

w x8 E. Rich, K. Knight, Artificial Intelligence, McGraw-Hill,

New York, 1991, pp. 487509.w x9 D.J. Roe, M. Cheng, V.A. Elberry, P.E. Nolan, Estimating

parameters in non-linear mixed effects models: a comparison

using clinical data in children, Proceedings of the Biophar- .maceutical Section, American Statistical Association 1994

318.w x10 D. Rummelhart, R. Durbin, R. Golden, Y. Chauvin, Back-

propagation: the basic theory, Backpropagation Theory, Ar- .chitectures, and Applications 1995 134.

w x11 D. Rummelhart, B. Widrow, M. Lehr, The basic ideas in . .neural networks, Communications of the ACM 37 3 1994

8792.w x12 R. Schaff, G. Wasserman, R. Englebrecht, W. Scholz, Medi-

cal treatment assistance with an interactive drug information

system, Medical Expert Systems Using Personal Computers .1987 4551.

w x13 L.B. Sheiner, T.M. Ludden, Population pharmacokineticsr

dynamics, Annual Review of Pharmacological and Toxicol- .ogy 32 1992 185209.

Kristin M. Tolle is a PhD candidate of

Management Information Systems at theUniversity of Arizona where she re-

.ceived her MS in MIS 1997 . She is

also a senior member and research asso-

ciate of the UArMIS Artificial Intelli-

gence Lab. She received her BS in

Computer Information Systems from .Boise State University 1988 . Tolle is a

recipient of a research fellowship from

the National Library of Medicine and

the Oak Ridge Institute for Science and

Education. She has published several journal and conference

articles on topics ranging from medical information retrieval,

natural language processing, intelligent agents and neural networkmedical decision support systems.

Dr. Hsinchun Chen is a McClelland Pro-

fessor of Management Information Sys-

tems at the University of Arizona and

head of the UArMIS Artificial Intelli-

gence Lab. He received the PhD degree

in Information Systems from New York

University in 1989. He is the author of

more than 70 journal articles covering

semantic retrieval, search algorithms,

knowledge discovery and collaborative

computing in leading information tech-

nology publications such as DecisionSupport Systems, Journal of the American Society for Information

Science, and Communications of the ACM. He serves on the

editorial board of the Journal of the American Society for Infor-

mation Science and Decision Support Systems. He is an expert in

digital library and knowledge management research, whose work

has been featured in various scientific and information technolo-

gies publications including Science, Business Week, NCSA Ac-

cess Magazine, WEBster and HPCWire.


13/13


Dr. Hsiao-Hui Chow is an Assistant

Professor of Pharmacy at the University

of Arizona and an affiliate member of

the Arizona Cancer Center. Dr. Chow

received her bachelors of Science de- .gree from Taipei Medical College 1983

and her PhD from State University of .New York at Buffalo 1989 . She con-

ducts research on the characterization of

the disposition of drugs in the body,

identification and understanding factors

which may influence the disposition ki-

netics and responses of xenobiotics, and development of pharma-

cokinetic and pharmacodynamic models in quantitating and pre-

dicting the kinetic processes of drug absorption, distribution,

elimination and response. She has published several articles in the

areas of pharmacokinetics and biopharmaceutics.

ANN Blood Plasma Level

Documents

Transcript of ANN Blood Plasma Level